Migrate from GitLab to GitHub with Docker packaging

[csparker247]

Summary

Migrates spectral-analysis from GitLab to GitHub: ports CI to GitHub Actions, adds a multi-arch Docker image published to GHCR (with PR-time build validation), retires the Singularity definition in favor of that image, and raises the package to Python 3.12+ to match what the dependency stack now permits.

Changes

Workflows and CI

• Add .github/workflows/ci.yml: matrix across Python 3.12 and 3.13; installs ExifTool via apt (libimage-exiftool-perl, not the bare exiftool package, which only exists on Debian), runs pip install ., smoke-tests all three console scripts, then runs the unittest suite. Triggers on main pushes, v* tags, PRs, and manual dispatch.
• Add .github/workflows/build_docker.yml: builds the root Dockerfile and pushes to ghcr.io/educelab/spectral-analysis. Builds linux/amd64,linux/arm64 (arm64 under QEMU emulation) everywhere. On main/tags it publishes; on pull requests it builds the same platforms but does not push, as a Dockerfile-breakage check — so a PR validates exactly what a publish will build. Registry login is skipped on PRs. Tags: latest (release tags only), edge (main tip), and semver tags (e.g. 3.0.0).
• Concurrency grouped per-ref, with cancel-in-progress only for pull requests, so PR validation builds neither serialize behind each other nor behind a release publish.
• All actions pinned to current majors, off the deprecated Node 20 runtime: checkout v7, setup-python v7, metadata-action v6, login-action v4, setup-qemu-action v4, setup-buildx-action v4, build-push-action v7.
• Remove .gitlab-ci.yml, replaced entirely by the above.

Docker image

• Add Dockerfile: python:3.13-slim base, chosen because every compiled dependency (numpy, scipy, imagecodecs, scikit-image, scikit-learn, pillow) has manylinux wheels for both amd64 and arm64 there, so the image needs no build toolchain. Installs ExifTool from source with SourceForge as the primary mirror and a gzip -t validation guard; installs the package non-editable; sets WORKDIR /data; CMD defaults to spec-enhance -h.
• Add .dockerignore: excludes .git, __pycache__, legacy/, local venvs, test data, and built containers.

Python support

• Raise python_requires to >=3.12, and update the CI matrix and README requirements to match.

Documentation and configuration

• README.md: fix pyhon3 typo, update clone URL, add Docker pull/run examples and an Apptainer section covering apptainer pull docker://ghcr.io/....
• CLAUDE.md: replace the GitLab CI section with GitHub Actions docs, record the Python floor and image base rationale, and replace the singularity/ description.
• setup.cfg: project URL to github.com; version 2.1.0 -> 3.0.0.

Retired

• Remove singularity/spectral-analysis.def and singularity/submit_job.sh. On the cluster, pull with apptainer pull docker://ghcr.io/educelab/spectral-analysis:latest. This drops the writable-overlay, edit-source-in-container workflow; submit_job.sh's SBATCH header and rclone bind settings remain recoverable from git history.

Notes / follow-ups

• Repository visibility: the repo must be made public before the first v* tag push, or the published GHCR package will be private.
• Python 3.12+ is a dependency constraint, not a preference: tifffile declares requires_python >=3.12 and numpy 2.5 dropped cp310/cp311 wheels, so older interpreters cannot install current versions of this project's own dependencies. Python 3.10 reaches EOL in October 2026; 3.11 in October 2027.
• Version number: bumped to 3.0.0. Raising python_requires breaks installation for 3.10/3.11 users, so this is a major bump rather than a minor one. A larger refactor is expected before the release is published, so this may move again. ENHANCE_VERSION in enhance.py is versioned independently and was left alone, since spec-enhance's behavior is unchanged.
• Docker build does not gate on tests: build_docker.yml runs independently of ci.yml, mirroring the sibling preppy repo. A red test suite can still publish a tagged image, including latest.
• PR builds cover both arches, so an arch-specific Dockerfile failure is caught before merge rather than after. The cost is wall-clock: amd64-only took 1m41s, both arches take ~13m45s, essentially all of it the emulated arm64 leg (ExifTool's make test runs its perl suite under QEMU, and the pip install is emulated too). If that becomes annoying, the levers are dropping make test from the ExifTool build, or moving to native ubuntu-24.04-arm runners — free for public repos, and no emulation at all.
• ExifTool is intentionally unpinned: exiftool.org only ever hosts the current tarball, so pinned URLs 404 once a new version ships (the reason 459929d moved the old .def to ver.txt). Distribution is really via SourceForge, so that is the primary source with exiftool.org as fallback, retries, and a gzip -t guard against the SF redirector returning an HTML error page. Mirrors pgs-recon's Dockerfile.
• Docker image is verified building on both arches: the image had never been built anywhere (no local Docker daemon), so the new pull_request trigger produced its first real builds on this PR. amd64 and arm64 both succeed, login is correctly skipped, and nothing is pushed. The remaining unexercised path is the publish itself, which cannot run until this merges to main.
• Local verification: pip install ., all three console-script smoke tests, and the unittest suite pass on Python 3.12 and 3.13; Python 3.10 is now correctly rejected by python_requires. A real spec-enhance run was verified end-to-end, including the ExifTool -Software= tag.
• CI results on this branch: the first push passed all four legs (3.10, 3.11, 3.12, 3.13) before the matrix was narrowed; the current matrix (3.12, 3.13) also passes.
• Prior GitLab CI bug, for the record: the old test:3.11 job specified image: python:3.10, so 3.11 was never actually tested despite CI reporting green on it. It is no longer tested here either — 3.11 is now below the supported floor.
• Test coverage: the suite is a single test asserting setup_logging sets the log level, so CI is an install-and-import check rather than a behavioral one. Tracked in Expand the test suite beyond the single smoke test #2 — worth resolving before any large refactor.
• GitLab issues: the 27 open issues intentionally remain on GitLab and were not migrated. All predate the 2.0 rewrite and reference the legacy/ package.

🤖 Generated with Claude Code